Semi-Supervised Natural Face De-Occlusion

Occlusions are often present in face images the wild, e.g., under video surveillance and forensic scenarios. Existing de-occlusion methods limited as they require knowledge of an occlusion mask. To overcome this limitation, we propose paper a new generative adversarial network (named OA-GAN) for natural without mask, enabled by learning semi-supervised fashion using (i) paired with known masks artificial occlusions (ii) masks. The generator our approach first predicts which is used filtering feature maps input image semantic cue de-occlusion. filtered then completion to recover non-occluded image. initial mask prediction might not be accurate enough, but it gradually converges one because loss use perceive regions need recovered. discriminator consists loss, distinguishing recovered from images, attribute preserving ensuring that after can retain attributes Experimental evaluations on widely CelebA dataset collected show proposed outperform state art